JP2002204175A - Method and apparatus for removing noise - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a method and an apparatus for removing noise capable of obtaining an emphasized voice with reduced distortion and noise regardless of types of noise and values of an SNR. SOLUTION: The apparatus has a weighted degraded-voice calculating section 14 for calculating a weighted degraded-voice power spectrum from a degraded voice power spectrum and an estimated noise power spectrum. Also, the apparatus has a suppression coefficient correcting section for calculating a corrected suppression coefficient in response to the value of the SNR and a suppression coefficient.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method and an apparatus for removing noise superimposed on a desired audio signal.

[0002]

2. Description of the Related Art A noise canceller is a technique for eliminating noise (noise) superimposed on a desired voice signal. The noise canceller estimates a power spectrum of a noise component using an input signal converted into a frequency domain. By subtracting the estimated power spectrum from the input signal, an operation is performed to suppress noise mixed in the desired audio signal. The power spectrum of the noise component can be applied to suppression of non-stationary noise by detecting and updating a silent section of a voice.

As a noise canceller, for example,
"December 1984, IEE Transactions on Axistics Speech and &
Signal Processing, Vol. 32, No. 6 (IEEE TRANSA
CTIONS ON ACOUSTICS, SPEECH, AND SIGNAL PROCESSING,
VOL.32, NO.6, PP.1109-1121, DEC, 1984), 1109-112
One page ”(Reference 1). This is known as the minimum mean square error short-time spectral amplitude method.

FIG. 24 shows the configuration of the noise canceller described in Reference 1.

[0005] A degraded audio signal (a signal in which a desired audio signal and noise are mixed) is supplied to an input terminal 11 as a sample value sequence. The degraded audio signal sample is supplied to the frame division unit 1 and is divided into frames every K / 2 samples. Here, K is an even number. The degraded audio signal sample divided into frames is supplied to the windowing processing unit 2 and multiplied by a window function w (t). Input signal y _n of the n-th frame (t) (t = 0, 1, ..., K / 2-1) with respect to w (t) signal window was morning by y _n (t) bar is the following formula Given.

[0006]

(Equation 1)

It is also widely practiced to overlap (overlap) a part of two consecutive frames to form a window. Assuming that 50% of the frame length is used as the overlap length, y _n (t) bar (t
= 0, 1,..., K−1) are outputs of the windowing processing unit 2.

[0008]

(Equation 2)

Hereinafter, description will be given of a case where windowing is performed by overlapping 50% of two consecutive frames. w (t)
For example, a Hanning window shown in Expression (4) can be used.

[0010]

(Equation 3)

The windowed output y _n (t) bar is supplied to a Fourier transform unit 3 and converted into a degraded speech spectrum Y _n (k). The degraded voice spectrum Y _n (k) is separated into a phase and an amplitude, the arg Y _n (k) of the degraded voice phase spectrum is supplied to the inverse Fourier transform unit 9, and the degraded voice amplitude spectrum | Y _n (k) | 4. It is supplied to the multiplex multiplier 16 and the multiplex multiplier 17.

[0012] The voice detection unit 4 provides a degraded voice amplitude spectrum.
The presence or absence of speech is detected based on | Y _n (k) |, and a speech detection flag determined by the result is transmitted to estimated noise calculation section 51. The multiplexing unit 17 calculates the degraded speech power spectrum using the supplied degraded speech amplitude spectrum | Y _n (k) |, and sends it to the estimated noise calculation unit 51 and the SNR (signal-to-noise ratio) calculation unit 6 for each frequency. introduce.

The estimated noise calculation unit 51 estimates the power spectrum of the noise using the voice detection flag, the degraded voice power spectrum, and the count value supplied from the counter 13, and obtains the frequency-dependent SNR calculation unit 6 as the estimated noise power spectrum. To communicate. The frequency-specific SNR calculator 6 calculates the SNR for each frequency using the input degraded voice power spectrum and the estimated noise power spectrum, and supplies the calculated SNR to the estimated a priori SNR calculator 7 and the noise suppression coefficient generator 8 as an acquired SNR. I do.

The estimated a priori SNR calculation unit 7 estimates an a priori SNR using the acquired a posteriori SNR and the suppression coefficient supplied from the noise suppression coefficient generation unit 8, and performs noise suppression as the estimated a priori SNR. It returns to the coefficient generator 8.

The noise suppression coefficient generation unit 8 generates a noise suppression coefficient using the acquired SNR and the estimated innate SNR supplied as inputs, and feeds the noise suppression coefficient back to the estimated innate SNR calculation unit 7 as a suppression coefficient. The information is transmitted to the unit 16.

The multiplex multiplier 16 is supplied from the Fourier transformer 3
Degraded speech amplitude spectrum | Y _n(k) |
Suppression coefficient G supplied from coefficient generation unit 8_n(k) Weight with bar
Speech amplitude spectrum | X_n(k) |
The bar is obtained and transmitted to the inverse Fourier transform unit 9. | X_n(k) |
The bar is given by equation (5).

[0017]

(Equation 4)

The inverse Fourier transform unit 9 includes the emphasized speech amplitude spectrum | X _n (k) | supplied from the multiplex multiplier 16 and the degraded speech phase spectrum a supplied from the Fourier transform unit 3.
Multiply by rg Y _n (k) to obtain an enhanced voice X _n (k) bar. That is, equation (6) is executed.

[0019]

(Equation 5)

The obtained emphasized voice X _n (k) bar is subjected to inverse Fourier transform, and a time-domain sample value sequence x _n (t) bar (t = 0, 1,...) In which one frame is composed of K samples. ., K-1) to the frame synthesizing unit 10. Frame synthesis unit 10
Extracts K / 2 samples from two adjacent frames of the x _n (t) bar and superimposes them, and obtains the emphasized voice x _n (t) hat by Expression (7). X _n (t) hat obtained
(t = 0, 1,..., K−1) are transmitted to the output terminal 12 as the output of the frame synthesis unit 10.

[0021]

(Equation 6)

Document 1 does not disclose in detail the method of realizing the voice detection unit. However, as an example of the realization of the voice detection unit, "March 2000, Proceedings of the Acoustical Society of Japan,
322 page (Reference 2) is known.
Will be described as a conventional method.

FIG. 25 is a block diagram showing a configuration of the voice detection unit 4 included in FIG. The voice detection unit 4 includes a threshold storage unit 401, a comparison unit 402, a multiplier 404, a logarithmic calculation unit 405, a power calculation unit 406, a weighted addition unit 407, a weight storage unit 408, and a logical NOT circuit 409.

The degraded voice amplitude spectrum supplied from the Fourier transform unit 3 in FIG. 24 is supplied to the power calculation unit 406. The power calculator 406 calculates the sum of the power | Y _n (k) | ² of the degraded voice amplitude spectrum from k = 0 to K−1,
The result is transmitted to the logarithmic calculation unit 405. The logarithmic calculation unit 405 calculates the logarithm of the input degraded voice spectrum power, and
To communicate. The multiplier 404 obtains a noisy speech power Q _n by a constant multiple of the supplied logarithm, and supplies this to the comparing unit 402 and the weighted adder 407. That is, the degraded voice power Qn of the _n- th frame is given by the following equation.

[0025]

(Equation 7)

Note that the voice detection unit disclosed in Reference 2
Using y _n (t) bar, which is a time domain sample, equation (9)
Seeking a Q _n accordance with.

[0027]

(Equation 8)

However, as shown in, for example, "1985, Theory of Digital Signal Processing, Corona, pp. 75-76" (Reference 3), the equation (9) and the equation (8) are equivalent. Also known as the Parseval equation.

The comparison unit 402 stores the threshold value from the threshold value storage unit 401
TH _n is supplied. Comparing unit 402 compares the output with the threshold value TH _n multiplier 404, a representative of the sound "1" when the TH _n> Q _n, a "0" representing the silence when the TH _n ≦ Q _n Is output as a voice detection flag. The output of the comparison unit 402 is supplied to the outside as a voice detection flag, which is the output of the voice detection unit 4, and at the same time, is supplied to the negative operation circuit 409. The output of the NOT operation circuit 409 is supplied to the weighted addition unit 407 as a weighted addition unit control signal 905. Weighted adder 4
07 also includes a threshold from the threshold storage unit 401 and a weight storage unit 4
08 is supplied with weights.

The weighted addition section 407 selectively updates the threshold 902 supplied from the threshold storage section 401 based on the weighted addition section control signal 905, and sets the updated threshold 904 as the threshold storage section 4.
Return to 01. Updating the threshold value TH _n is the threshold value TH _n-1 and noisy speech power 901 is determined by adding weighted using the weight 903 supplied from the weight storage unit 408. Update threshold
TH _n is calculated only when the weighted adder control signal 905 output from the logical NOT circuit 409 is equal to “1”. That is, only when the silent threshold TH _n is updated. The update threshold 904 obtained by the update is stored in the threshold storage unit 4
Returned to 01.

FIG. 26 shows a power calculation unit 406 included in FIG.
FIG. 3 is a block diagram showing the configuration of FIG. The power calculation unit 406
It has a separation unit 4061, multipliers 4062 _{0 to} 4062 _K−1 , and an adder 4063. The degraded voice amplitude spectrum supplied from the Fourier transform unit 3 in FIG. 24 in a multiplexed state is separated into K samples for each frequency by a separating unit 4061,
Is supplied to the 2 ₀ ~4062 _K-1. Multipliers 4062 _{0 to} 4062 _K-1 are:
Each input signal is squared and transmitted to the adder 4063. The adder 4063 calculates and outputs the sum of the input signals.

FIG. 27 shows a weighted adding section 40 included in FIG.
FIG. 9 is a block diagram showing the configuration of FIG. Weighted adder 407
Are multipliers 4071, 4073, constant multiplier 4075, adder 407
2,4074.

The degraded audio power 901 is output from the multiplier 404 in FIG.
However, a threshold 902 from the threshold storage unit 401 in FIG. 25, a weight 903 from the weight storage unit 408 in FIG. 25, and a weighted addition unit control signal 905 from the logical NOT circuit 409 in FIG. 25 are supplied as inputs. The weight 903 having the value β is transmitted to the constant multiplier 4075 and the multiplier 4073. The constant multiplier 4075 converts the input signal to -1
The obtained −β is transmitted to the adder 4074. Adder 4
0 is supplied as the other input of 074, and the output of the adder 4074 is 1-β which is the sum of the two. 1-beta is supplied to the multiplier 4071 is multiplied by the noisy speech power Q _n which is the other input is the product (1-beta) Q _n adder 4
072. On the other hand, in the multiplier 4073, β supplied as the weight 903 and the threshold 902 are multiplied, and the product β TH
_n-1 is transmitted to the adder 4072. The adder 4072 calculates βTH _n-1
And the sum of (1−β) Q _n is output as the update threshold value 904.
Calculation of updating the threshold TH _n is weighted adder control signal 905 is performed only when equal to "1". That is, the function of weighted adder, when silence is that obtaining the TH _n to update the threshold value TH _n-1, can be represented by the following equation. In the following equation, β is the value of the weight 903.

[0034]

(Equation 9)

FIG. 28 is a block diagram showing a configuration of the multiplexing / multiplying unit 17 included in FIG. The multiplexing / multiplying unit 17 includes multipliers 17010 _{0 to} 1701 _K−1 , separating units 1702 and 1703, and a multiplexing unit 1704. The degraded voice amplitude spectrum supplied from the Fourier transform unit 3 in FIG. 24 in a multiplexed state is separated into K samples for each frequency in separation units 1702 and 1703, and supplied to multipliers 17010 _{0 to} 1701 _K-1 , respectively. Is done. Multiplier 1701 ₀
1701701 _K−1 squares the input signals and transmits them to the multiplexing unit 1704. The multiplexing unit 1704 multiplexes the input signal and outputs the multiplexed signal as a degraded audio power spectrum.

FIG. 29 shows an estimated noise calculator 51 included in FIG.
FIG. 3 is a block diagram showing the configuration of FIG. The estimation noise calculation unit 51
Separation unit 502, a multiplexing unit 503, a frequency domain estimated noise calculator 514 ₀
~ 514 _K-1 . The voice detection flag supplied from the voice detection unit 4 in FIG. 24 and the count value supplied from the counter 13 in FIG. 24 are transmitted to the frequency _- dependent estimated noise calculation units 514 _{0 to} 514 _K−1 . The degraded voice power spectrum supplied from the multiplex multiplier 17 in FIG. 24 is transmitted to the demultiplexer 502. Separating section 502 separates the degraded voice power spectrum supplied in a multiplexed state into components corresponding to K frequencies, and transmits the components to frequency _- dependent estimated noise calculating sections 514 _{0 to} 514 _K−1 .
The frequency _- dependent estimated noise calculators 514 _{0 to} 514 _K−1 calculate the noise power spectrum using the degraded voice power spectrum supplied from the separator 502 and transmit the noise power spectrum to the multiplexer 503. The calculation of the noise power spectrum is controlled by the count value and the value of the voice detection flag, and is executed only when a predetermined condition is satisfied. The multiplexing unit 503 determines the supplied K
The noise power spectrum values are multiplexed and output as an estimated noise power spectrum.

FIG. 30 is a block diagram showing a configuration of frequency-based estimated noise calculation section 514 included in FIG. The noise estimation disclosed in Reference 2 updates a noise estimation value in a silent section, and uses an instantaneous value of estimated noise averaged by a recursive filter as the noise estimation value. Meanwhile, `` May 1998, IEE Transactions on Speech and Audio Processing, Volume 6, Issue 3 (IEEE TRANSACTIONS ON SPEE
CH AND AUDIO PROCESSING, VOL. 6, NO. 3, PP.287-292,
MAY, 1998), pp. 287-292 ”(Reference 4) describes that the instantaneous value of the estimated noise is averaged and used. This suggests the realization of averaging using a transversal filter (a configuration using a shift register) instead of the cyclic filter. Since the functions are the same in both implementations, the method disclosed in Reference 4 will be described here.

The frequency-dependent estimated noise calculation section 514 includes an update determination section 521, a register length storage section 5041, a switch 5044, a shift register 5045, an adder 5046, a minimum value selection section 5047, and a division section.
5048 and a counter 5049. Switch 5044 has a
The degraded voice power spectrum for each frequency is supplied from the separation unit 502. When the switch 5044 closes the circuit, the degraded audio power spectrum for each frequency is transmitted to the shift register 5045. The shift register 5045 shifts the value stored in the internal register to an adjacent register according to the control signal supplied from the update determination unit 521. The shift register length is equal to a value stored in a register length storage unit 5941 described later. All register outputs of the shift register 5045 are supplied to an adder 5046. The adder 5046 adds all the supplied register outputs, and transmits the addition result to the divider 5048.

On the other hand, the update determination unit 521 is supplied with a count value and a voice detection flag. The update determination unit 521
Until the count value reaches a preset value, the count value is always "1". After the count value reaches "1" when the voice detection flag is "0" (silence), otherwise "0". Is output to the counter 5049, the switch 5044, and the shift register 5045. The switch 5044 closes the circuit when the signal supplied from the update determination unit is “1” and opens the circuit when the signal is “0”. The counter 5049 increases the count value when the signal supplied from the update determination unit is “1”, and does not change the count value when the signal is “0”. The shift register 5045 captures one sample of the signal sample supplied from the switch 5044 when the signal supplied from the update determination unit is “1”, and simultaneously shifts the stored value of the internal register to the adjacent register. The output of the counter 5049 and the output of the register length storage unit 5941 are supplied to the minimum value selection unit 5047. The minimum value selection unit 5047 selects one of the supplied count value and register length.
The smaller one is selected and transmitted to the division unit 5048. Divider 50
48 divides the sum of the frequency-dependent degraded audio power spectrum supplied from the adder 5046 by the smaller value of the count value or the register length, and outputs the quotient as the frequency-dependent estimated noise power spectrum λ _n (k) . B _n (k) (n = 0, 1, ..., N
Assuming that -1) is a sample value of the degraded sound power spectrum stored in the shift register 5045, λ _n (k) is given by Expression (11).

[0040]

(Equation 10)

Here, N is the smaller value of the count value and the register length. Since the count value monotonically increases starting from zero, division is first performed by the count value, and then division is performed by the register length. When the division is performed by the register length, an average value of the values stored in the shift register is obtained. At first, not enough values are stored in the shift register 5045,
Divide by the number of registers where the value is actually stored. The number of registers in which the values are actually stored is equal to the count value when the count value is smaller than the register length, and equal to the register length when the count value is larger than the register length.

FIG. 31 is a block diagram showing a configuration of the update determination unit 521 included in FIG. The update determination unit 521 includes a logical NOT circuit 5202, a comparison unit 5203, a threshold storage unit 5204, and a logical sum calculation unit 5211. The count value supplied from counter 13 in FIG. 24 is transmitted to comparing section 5203. Threshold storage unit 5204
Is also transmitted to the comparing section 5203. Comparison section
The 5203 compares the supplied count value with the threshold value, and transmits "1" to the logical sum calculation unit 5211 when the count value is smaller than the threshold value and "0" when the count value is larger than the threshold value. On the other hand, the supplied voice detection flag is a logical NOT circuit
Transferred to 5202. The logical NOT circuit 5202 obtains a logical NOT value of the input signal and transmits it to the logical sum calculating unit 5211.
In other words, “0” is transmitted to the sound portion where the voice detection flag is “1”, and “1” is transmitted to the logical sum calculation portion 5211 in the silent portion where the voice detection flag is “0”. As a result, the output of the logical sum calculation unit 5211 becomes “1” when the sound detection flag is a silent part with “0” or when the count value is smaller than the threshold value, and closes the switch in FIG.
The counter 5049 is counted up.

FIG. 32 is a block diagram showing the configuration of the frequency-specific SNR calculator 6 included in FIG. Frequency domain SNR calculator 6, the division unit 601 ₀ ~601 _K-1, the separation unit 602 and 603 includes a multiplexer 604. The degraded audio power spectrum supplied from the multiplex multiplier 17 in FIG. 24 is transmitted to the demultiplexer 602. The estimated noise power spectrum supplied from the estimated noise calculation unit 51 in FIG. 24 is transmitted to the separation unit 603. The degraded voice power spectrum is separated into K samples corresponding to the frequency components in the separation unit 602, and the estimated noise power spectrum is separated in the separation unit 603.
Supplied from _{0 to} 601 _K-1 . The divider 601 ₀ ~601 _K-1, wherein
According to (12), the supplied degraded voice power spectrum is divided by the estimated noise power spectrum to obtain a frequency-dependent SNRγ.
_n (k) is obtained and transmitted to the multiplexing unit 604.

[0044]

[Equation 11]

Where λ _n (k) is the estimated noise power spectrum. The multiplexing unit 604 transmits the transmitted K frequency-specific SNs.
R is multiplexed and output as an acquired SNR.

FIG. 33 is a block diagram showing the configuration of the estimated innate SNR calculator 7 included in FIG. Estimated innate SNR
The calculation unit 7 includes a multi-value range limitation processing unit 701, an acquired SNR storage unit
702, a suppression coefficient storage unit 703, multiplex multiplication units 704 and 705, a weight storage unit 706, a multi-weighted addition unit 707, and an adder 708.

The acquired SNRγ _n (k) (k = 0, 1,..., K−1) supplied from the frequency-specific SNR calculation unit 6 in FIG. 24 is obtained by the acquired SNR storage unit 702 and the adder 708. Is transmitted to Acquired SNR storage unit 702
May store the acquired SNRganma _n (k) in the n-th frame, and transmits acquired in the n-1 frame SNRγn-1 (k) of the multiplexed multiplier 705. The suppression coefficient G _n (k) bar (k = 0, 1,..., K−1) supplied from the noise suppression coefficient generation unit 8 in FIG.
Is transmitted to the suppression coefficient storage unit 703. Suppression coefficient storage unit 7
03 stores the suppression coefficient G _n (k) bar in the n-th frame and transmits the suppression coefficient G _n-1 (k) bar in the n−1-th frame to the multiplexing unit 704. The multiplex multiplier 704 squares the supplied G _n-1 (k) bar to obtain G ² _n-1 (k) bar, and transmits it to the multiplex multiplier 705. The multiplex multiplication unit 705 multiplies G = 0 (1, ..., K−1) by G ² _n−1 (k) bar and γ _n−1 (k) to obtain G
² _n−1 (k) bar γ _n−1 (k) is obtained, and the result is transmitted to the multi-weighted addition unit 707 as the past estimated SNR 922. The configuration of the multiplex multipliers 704 and 705 in FIG. 33 is the same as the configuration of the multiplex multiplier 17 already described with reference to FIG. 28, and thus detailed description will be omitted.

The other terminal of the adder 708 is supplied with −1, and the addition result γ _n (k) −1 is transmitted to the multi-value range limiting processing unit 701. The multi-value range limiting processing unit 701 performs an operation using the value-limited operator P [•] on the addition result γ _n (k) -1 supplied from the adder 708, and obtains a result P [γ _n (k) -1. ] To the multi-weighted addition section 707 as the instantaneous estimation SNR 921. Here, P [x] is determined by Expression (13).

[0049]

(Equation 12)

The weight 923 is supplied from the weight storage unit 706 to the multiple weighted addition unit 707. The multi-weighted adding unit 707 uses these supplied instantaneous estimated SNR 921, past estimated SNR 922, and weight 923 to estimate an innate SNR 924.
Ask for. Assuming that the weight 923 is α and the ξ _n (k) hat is an estimated innate SNR, the ξ _n (k) hat is calculated by Expression (14).

[0051]

(Equation 13)

Here, it is assumed that G ² ₋₁ (k) γ ₋₁ (k) bar = 1.

FIG. 34 is a block diagram showing a configuration of multiplexed range limiting processing section 701 included in FIG. The multiple value range limitation processing unit 701 includes a constant storage unit 7011, maximum value selection units 7012 _{0 to} 7012.
_K-1 , a demultiplexing unit 7013 and a multiplexing unit 7014. Separation unit 7013
Γ _n (k) −1 is supplied from the adder 708 in FIG.
Separating section 7013 separates supplied γ _n (k) -1 into K frequency _- dependent components, and supplies the result to maximum value selecting sections 7012 _{0 to} 7012 _K-1 .
Zeros are supplied from the constant storage unit 7011 to the other inputs of the maximum value selection units 70120 to 7012K-1. Maximum value selection section 7012
_{For 0 to} 7012 _K-1 , γ _n (k) -1 is compared with zero, and the larger value is transmitted to the multiplexing unit 7014. This maximum value selection operation is based on the expression
This corresponds to executing (13). The multiplexing unit 7014 multiplexes these values and outputs them.

FIG. 35 is a block diagram showing a configuration of the multiple weighted addition section 7071 included in FIG. The multiple weighted addition unit 7071 includes weighted addition units 7071 _{0 to} 7071 _K−1 , a separation unit 70
72, 7074 and a multiplexing unit 7075. Figure 70
P [γ _n (k) −1] is supplied as the instantaneous estimation SNR 921 from the multiple value range limitation processing section 701 of 33. The separating unit 7072 calculates P [γ _n
(k) -1] is separated into K frequency-specific components, and weighted addition units 7071 _{0 to} 7071 are obtained as frequency-specific instantaneous estimation SNRs 921 _{0 to} 921 _K-1.
Transmit to _K-1 . Demultiplexing section 7074 includes multiplex multiplying section 705 in FIG.
, The G ² _n−1 (k) bar γ _n−1 (k) is supplied as the past estimated SNR 922. The separation unit 7074 has a G ² _n-1 (k) bar γ _n-1 (k)
Is separated into K frequency components, and the past estimated SNR for each frequency
922 _{0 to} 922 _K-1 , weighted adder 7071 _{0 to} 7071 _K-1
To communicate. On the other hand, the weighted adder 7071 ₀ ~7071 _K-1 is the weight 923 is also supplied. Weighted adder 7071 _{0 to} 707
1 _K-1 performs weighted addition represented by equation (14), and multiplexes the estimated innate SNRs 924 _{0 to} 924 _K-1 for each frequency by the multiplexing unit 70.
Communicate to 75. The multiplexing unit 7075 calculates the estimated innate SN for each frequency.
R 924 _{0 to} 924 _K-1 are multiplexed and output as estimated innate SNR 924.

The operation and configuration of the weighted adders 7071 ₀ to 7071 _K-1 are the same as those of the weighted adder 40 already described with reference to FIG.
Since it is equal to 7, detailed description is omitted. However, the calculation of the weighted addition is always performed.

FIG. 36 is a block diagram showing the configuration of the noise suppression coefficient generator 8 included in FIG. Spectral gain generator 8, the suppression coefficient search unit 801 ₀ ~801 _K-1, the separation section 802,80
3. It has a multiplexing unit 804. The acquired SNR is supplied to the separation unit 802 from the frequency-specific SNR calculation unit 6 in FIG. Separation unit 80
2, the supplied posteriori SNR separated into K frequency-component, and transmits the spectral gain search unit 801 ₀ ~801 _K-1. Separation unit 80
3 includes the estimated innate SNR from the estimated innate SNR calculation unit 7 in FIG.
R is supplied. The separation unit 803 determines the supplied estimated innate SN
Separating the R into K frequency-component suppression coefficient search unit 801 ₀ ~
Transmit to 801 _K-1 . Suppression coefficient search unit 801 ₀ ~801 _K-1 searches the suppression coefficient corresponding to the estimated apriori SNR and supplied posteriori SNR, to transmit the search result to the multiplexer 804. Multiplexing section 804 multiplexes the supplied suppression coefficients and outputs the result.

FIG. 37 shows the suppression coefficient search unit 80 included in FIG.
1 is a block diagram showing a _{0 ~801 K-1} configuration. The suppression coefficient search unit 801 includes a suppression coefficient table 8011 and an address conversion unit 801.
2,8013. The acquired SNR for each frequency is supplied to the address conversion unit 8012 from the separation unit 802 in FIG. The address conversion unit 8012 converts the obtained frequency-specific acquired SNR into a corresponding address and transmits the address to the suppression coefficient table 8011. The address conversion unit 8013 includes, from the separation unit 803 in FIG. 36,
An estimated innate SNR for each frequency is provided. Address converter 8
013 converts the supplied estimated innate SNR for each frequency into a corresponding address and transmits it to the suppression coefficient table 8011. The suppression coefficient table 8011 outputs the suppression coefficient stored in the area corresponding to the address supplied from the address conversion unit 8012 and the address conversion unit 8013 as the suppression coefficient for each frequency.

[0058]

In the conventional method described above, the power spectrum of the noise is updated in the silent section based on the output of the voice detection unit. For this reason, when the detection result of the voice detection unit is incorrect, the power spectrum of the noise cannot be accurately estimated. Even when a sound section lasts for a long time, the power spectrum of noise cannot be updated because there is no silent section, and the power spectrum estimation accuracy for non-stationary noise cannot be avoided. For this reason, there is a problem that noise and distortion remain in the emphasized voice.

In the conventional method, the power spectrum of noise is estimated using the degraded voice power spectrum. For this reason, under the influence of the voice power spectrum included in the degraded voice, the power spectrum of the noise cannot be accurately estimated, and the noise remains in the emphasized voice,
There is a problem that distortion occurs. Furthermore, in the conventional method, since noise suppression is performed for all SNRs using the suppression coefficient obtained by the same calculation method, there is a problem that it is not possible to achieve a sufficiently high enhanced speech sound quality.

An object of the present invention is to provide a noise removing method and apparatus capable of obtaining an emphasized speech with little distortion and noise by accurately estimating the power spectrum of the noise without being affected by the performance of the speech detection unit. It is to provide.

Another object of the present invention is to provide a noise elimination method capable of obtaining an emphasis voice with little distortion and noise with respect to non-stationary noise by accurately estimating the power spectrum of noise even in a sound section. It is to provide a method and an apparatus.

Another object of the present invention is to provide a noise removal method and apparatus capable of obtaining an emphasized voice with little distortion and noise by using an optimum suppression coefficient for every SNR value. is there.

[0063]

A method and apparatus for removing noise according to the present invention is characterized in that a power spectrum of noise is estimated using a weighted degraded speech power spectrum. More specifically, a weighted degraded speech calculation unit for calculating a weighted degraded speech power spectrum from the degraded speech power spectrum and the estimated noise power spectrum is provided.

The noise elimination method and apparatus of the present invention is characterized in that noise suppression is performed using a suppression coefficient corrected according to the value of SNR. More specifically, the present invention is characterized by including a suppression coefficient correction unit for receiving the SNR value and the suppression coefficient and calculating a corrected suppression coefficient.

[0065]

According to the present invention, the power spectrum of noise is estimated by using the weighted degraded voice power spectrum obtained from the degraded voice power spectrum and the estimated noise power spectrum. , And an enhanced voice with little distortion and noise can be obtained.

Further, in the present invention, noise suppression is performed using the suppression coefficient corrected according to the SNR value, so that it is possible to obtain an emphasized voice with little distortion and noise for all SNR values.

[0067]

FIG. 1 is a block diagram showing a first embodiment of the present invention. FIG. 1 is the same as FIG. 24, which is a block diagram of the conventional example, except for an estimated noise calculation unit 5, a weighted degraded speech calculation unit 14, and a suppression coefficient correction unit 15. Hereinafter, a detailed operation will be described focusing on these differences.

FIG. 2 is a block diagram showing the configuration of the weighted degraded speech calculation unit 14. The weighted degraded speech calculation unit 14
It has an estimated noise storage unit 1401, a frequency-dependent SNR calculation unit 1402, a multiplex nonlinear processing unit 1405, and a multiplex multiplication unit 1404.

The estimated noise storage section 1401 stores the estimated noise power spectrum supplied from the estimated noise calculation section 5 in FIG. 1, and outputs the estimated noise power spectrum stored one frame before to the SNR calculation section 1402 for each frequency. I do. The frequency-dependent SNR calculation unit 1402 obtains for each frequency using the estimated noise power spectrum supplied from the estimated noise storage unit 1401 and the SNR of the degraded voice power spectrum supplied from the multiplex multiplier 17 in FIG. Output to the processing unit 1405.

The multiplex nonlinear processing unit 1405 calculates a weight coefficient vector using the SNR supplied from the frequency-specific SNR calculation unit 1402, and outputs the weight coefficient vector to the multiplex multiplier 1404. The multiplex multiplication unit 1404 calculates, for each frequency, a product of the degraded voice power spectrum supplied from the multiplex multiplication unit 17 of FIG. 1 and a weight coefficient vector supplied from the multiplex nonlinear processing unit 1405, and obtains a weighted degraded voice power spectrum. Is output to the estimated noise storage unit 5 in FIG. SNR calculator for each frequency 140
The configuration of No. 2 is the same as that of the frequency-specific SNR calculation unit 6 already described with reference to FIG. Also, the configuration of the multiplex multiplier 1404 is the same as that of the multiplex multiplier 17 already described with reference to FIG. 28, and thus detailed description will be omitted.

Next, the configuration and operation of the multiplex nonlinear processing section 1405 of FIG. 2 will be described in detail with reference to FIG.
FIG. 3 is a block diagram showing a configuration of the multiplex nonlinear processing unit 1405 included in the weighted degraded speech calculation unit 14. The multiplex nonlinear processing unit 1405 includes a separating unit 1495, a nonlinear processing unit 1485 ₀ to
1485 _K-1 and a multiplexing unit 1475. The separation unit 1495
The SNR supplied from the SNR calculator for each frequency 1402 in FIG. 2 is separated into SNRs for each frequency, and the nonlinear processing units 1485 ₀ to 1485 _K−1
Output to Each of the non-linear processing units 1485 ₀ to 1485 _K-1 has a non-linear function that outputs a real value according to the input value. FIG. 4 shows an example of the nonlinear function. when the f ₁ and the input value, the output value f ₂ of the nonlinear function shown in FIG. 4, (1
It is given by equation 5).

[0072]

[Equation 14]

The non-linear processing units 1485 ₀ to 1485 _K−1 process the SNRs for each frequency supplied from the demultiplexing unit 1495 by using a non-linear function to obtain weight coefficients, and output the weighting coefficients to the multiplexing unit 1475. That is, the nonlinear processing units 1485 ₀ to 1485 _K-1
Outputs a weighting factor from 1 to 0 according to. Outputs 1 when the SNR is small, and outputs 0 when it is large. Multiplexer
The 1475 multiplexes the weighting factors output from the nonlinear processing units 1485 ₀ to 1485 _K-1 and multiplexes the weighting factor vector with the multiplexing unit.
Output to 1404.

The weighting factor multiplied by the degraded voice power spectrum in the multiplexing and multiplying unit 1404 in FIG. 2 has a value corresponding to the SNR. As the SNR increases, that is, as the voice component included in the degraded voice increases, The value of the weight coefficient decreases. In general, the degraded speech power spectrum is used to update the estimated noise.However, by weighting the degraded speech power spectrum used for updating the estimated noise in accordance with the SNR, the speech component included in the degraded speech power spectrum is updated. Can be reduced, and more accurate noise estimation can be performed. Although an example in which a non-linear function is used for calculating the weighting coefficient has been described, it is also possible to use an SNR function expressed in another form such as a linear function or a high-order polynomial other than the non-linear function.

FIG. 5 is a block diagram showing a configuration of the estimated noise calculator 5 included in the first embodiment of the present invention.
The estimation noise calculation unit 51 shown in FIG. 29 is different from the estimation noise calculation unit 51 in that the separation unit 505 exists and the frequency-dependent estimation noise calculation units 514 ₀ to 514.
Are identical except _{that K-1} has been replaced with a frequency domain estimated noise calculator 514 ₀ ~ 514 _K-1. Hereinafter, a detailed operation will be described focusing on these differences.

The separation unit 505 separates the weighted degraded speech power spectrum supplied from the weighted degraded speech calculation unit in FIG. 1 into the weighted degraded speech power spectrum for each frequency, and estimates the frequency-dependent estimated noise calculation unit 504 _0. ~ 504 Output to _K-1 . The frequency-dependent estimated noise calculation units 504 ₀ to 504 _K−1 are provided for the frequency-dependent degraded voice power spectrum supplied from the demultiplexing unit 502, the frequency-weighted degraded voice power spectrum supplied from the demultiplexing unit 505, and the voice detection of FIG. The estimated noise power spectrum for each frequency is calculated from the voice detection flag supplied from the unit 4 and the count value supplied from the counter 13 in FIG. Multiplexer 503
Multiplexes the estimated noise power spectrum for each frequency supplied from the estimated noise calculation section for each frequency 504 ₀ to 504 _K−1 ,
The estimated noise power spectrum is calculated by the frequency-dependent SNR calculator 6 in FIG.
Is output to the weighted degraded speech calculation unit 14. A detailed description of the configuration and operation of the frequency-dependent estimated noise calculators 504 ₀ to 504 _K−1 will be made with reference to FIG.

[0077] Figure 6 is a block diagram showing a frequency different estimated noise calculator 504 ₀ ~ 504 _K-1 configuration included in FIG.
The difference from the frequency-based estimated noise calculation section 514 shown in FIG. 30 is that the frequency-based estimated noise calculation sections 504 ₀ to 504 _K-1 have the estimated noise storage section 5942, and the update determination section 521 has the update determination section 520. And the switch 5044
Is replaced with the frequency-dependent weighted degraded voice power spectrum from the frequency-dependent degraded voice power spectrum. Estimated noise calculation unit for each frequency 504 ₀ to 504
_K-1 uses the weighted degraded voice power spectrum instead of the degraded voice power spectrum for the calculation of the estimated noise, and uses the estimated noise and the degraded voice power spectrum to determine whether to update the estimated noise. Differences occur. The estimated noise storage unit 5942 stores the estimated noise power spectrum for each frequency supplied from the division unit 5048,
The estimated noise power spectrum for each frequency stored before the frame is output to update determination section 520.

FIG. 7 is a block diagram showing a configuration of update determination section 520 included in FIG. Update determination unit 52 shown in FIG. 31
The difference from 1 is that the logical sum calculation unit 5211 is
And the update determination unit 520
5. It has a threshold storage unit 5206 and a threshold calculation unit 5207. Hereinafter, a detailed operation will be described focusing on these differences.

The threshold calculation unit 5207 calculates a value corresponding to the estimated noise power spectrum for each frequency supplied from the estimated noise storage unit 5942 in FIG.
Output to The simplest method of calculating the threshold is a constant times the estimated noise power spectrum for each frequency. In addition, it is also possible to calculate the threshold value using a higher-order polynomial or a nonlinear function. The threshold storage unit 5206 stores the threshold output from the threshold calculation unit 5207, and outputs the threshold stored one frame before to the comparison unit 5205. The comparison unit 5205 compares the threshold value supplied from the threshold value storage unit 5206 with the frequency-dependent degraded audio power spectrum supplied from the separation unit 502 in FIG. 5, and if the frequency-specific degraded audio power spectrum is smaller than the threshold value, “1”. Or, if larger, “0”.
Output to 01. That is, it is determined whether or not the degraded speech signal is noise based on the magnitude of the estimated noise power spectrum. The logical sum calculation unit 5201 includes an output value of the comparison unit 5203, an output value of the logical negation circuit 5202, and a comparison unit 520.
5 and outputs the calculation result to the switch 5044, the shift register 5045, and the counter 5049 in FIG.

As described above, when the deteriorated voice power is small not only in the initial state and the silent section, but also in the sound section, the update determination section 520 outputs “1”. That is, the estimation noise is updated. Since the calculation of the threshold is performed for each frequency, the estimation noise can be updated for each frequency.

In FIG. 6, CNT is the count value of the counter 5049, and N is the register length of the shift register 5045. And B _n (k) (n = 0,1, ..., N-1)
The frequency-dependent weighted degraded voice power spectrum stored in 5045 is used. At this time, the frequency-dependent estimated noise power spectrum λ _n (k) output from the division unit 5048 is given by Expression (16).

[0082]

(Equation 15)

That is, λ _n (k) is the shift register 5045
Is the average value of the frequency-dependent weighted degraded voice power spectrum stored in. The calculation of the average value can be performed using a weighted addition unit (recursive filter). Next, a configuration example in which a weighted addition unit is used for calculating λ _n (k) will be described with reference to FIG.

FIG. 8 is a block diagram showing a second configuration example of the frequency-dependent estimated noise calculation sections 504 ₀ to 504 _K-1 included in FIG. In the estimated noise calculating section 504 for each frequency in FIG. 6,
Shift register 5045, adder 5046, minimum value selector 5047,
Instead of the division unit 5048, the counter 5049, and the register length storage unit 5941, the frequency-based estimated noise calculation unit 507 includes a weighted addition unit 5071 and a weight storage unit 5072.

The weighted addition section 5071 is provided with an estimated noise storage section 59
Using the estimated noise power spectrum by frequency one frame before supplied from 42, the weighted degraded voice power spectrum by frequency supplied from switch 5044, and the weight output from weight storage unit 5072, the estimated noise by frequency is calculated. Then, output to multiplexing section 503. That is, the weight storage unit 507
When the weight stored in 2 is δ and the weighted degraded voice power spectrum by frequency is | Y _n (k) | ² bar, the estimated noise power spectrum by frequency λ _n (k) output from the weighted addition section 5071 Is given by equation (17). The configuration of weighted addition section 5071 is the same as that of weighted addition section 407 described with reference to FIG. 27, and a detailed description thereof will be omitted. However, the calculation of the weighted addition is always performed.

[0086]

(Equation 16)

FIG. 9 is a block diagram showing a configuration of the suppression coefficient correction unit 15 included in the first embodiment of the present invention.
When SNR is low, residual noise caused by insufficient suppression,
In order to prevent sound quality deterioration due to voice distortion generated by excessive suppression when the NR is high, the suppression coefficient correction unit 15 corrects the suppression coefficient according to the SNR. As an example of correction, SNR
When SNR is low, a correction value is added to the suppression coefficient to suppress residual noise, and when SNR is high, a lower limit value is set for the suppression coefficient to prevent voice distortion. Suppression coefficient correction unit 1
5 includes frequency-dependent suppression coefficient correction units 15011 ₁ to 1501 _K−1 , separation units 1502 and 1503, and a multiplexing unit 1504.

[0088] separating unit 1502 separates the estimated apriori SNR supplied from estimated apriori SNR calculator 7 of Figure 1 in frequency-components, frequency-suppression coefficients respectively corrector 1501 _0-1501
Output to _K-1 . The separation unit 1503 is the suppression coefficient generation unit of FIG.
The suppression coefficient supplied from 8 is separated into frequency-specific components and output to the frequency-dependent suppression coefficient correction units 15010 ₀ to 1501 _K-1 . Frequency-dependent suppression coefficient correction unit 1501 ₀ to 1501
_K-1 calculates a frequency-dependent corrected suppression coefficient from the estimated innate SNR for each frequency supplied from the demultiplexing section 1502 and the suppression coefficient for each frequency supplied from the demultiplexing section 1503, and
Output to 4. Multiplexing unit 1504 multiplexes the frequency-corrected suppression coefficient supplied from the frequency-suppression coefficient correction unit 1501 ₀ ~ 1501 _K-1, multiplexed multiplier 16 as corrected suppression coefficient
Is output to the estimated innate SNR calculator 7.

Next, the configuration and operation of the frequency-dependent suppression coefficient correction sections 15010 ₀ to 1501 _K-1 will be described in detail with reference to FIG. FIG. 10 is a block diagram showing a configuration of the frequency-dependent suppression coefficient correction units 15010 ₀ to 1501 _K-1 included in the suppression coefficient correction unit 15. The frequency-dependent suppression coefficient correction unit 1501
It has a maximum value selection unit 1591, a suppression coefficient lower limit storage unit 1592, a threshold storage unit 1593, a comparison unit 1594, a switch 1595, a correction value storage unit 1596, and a multiplier 1597.

The suppression coefficient lower limit storage section 1592 supplies the stored lower limit value of the suppression coefficient to the maximum value selection section 1591. The maximum value selection unit 1591 includes a suppression coefficient for each frequency and a suppression coefficient lower limit value storage unit 159 supplied from the separation unit 1503 in FIG.
The lower limit of the suppression coefficient supplied from 2 is compared, and the larger value is output to the switch 1595. That is, the suppression coefficient always has a value larger than the lower limit stored in the suppression coefficient lower limit storage unit 1592. Therefore, it is possible to prevent voice distortion caused by excessive suppression.

The comparison unit 1594 compares the threshold value supplied from the threshold value storage unit 1593 with the estimated innate SNR for each frequency supplied from the separation unit 1502 in FIG. 9, and if the estimated innate SNR for each frequency is larger than the threshold value. In this case, "0" is supplied to the switch 1595, and if it is smaller, "1" is supplied to the switch 1595. The switch 1595 outputs the signal supplied from the maximum value selection unit 1591 to the multiplier 1597 when the output value of the comparison unit 1594 is “1”.
Is output to the multiplexing unit 1504 in FIG. That is, when the estimated innate SNR for each frequency is smaller than the threshold, the suppression coefficient is corrected. By correcting the suppression coefficient when the SNR is small, it is possible to reduce the residual noise amount without excessively suppressing the voice component. The multiplier 1579 calculates the product of the output value of the switch 1595 and the output value of the correction value storage unit 1596, and divides the calculation result by using FIG.
To the multiplexing unit 1504. In order to reduce the suppression coefficient value, the correction value is usually smaller than 1, but this is not limited depending on the purpose. In the conventional example, the suppression coefficient is supplied to the multiplexing unit 16 and the estimated innate SNR calculation unit 7, but in the first embodiment of the present invention, the correction suppression coefficient is supplied instead of the suppression coefficient. .

FIG. 11 is a block diagram showing a second configuration example of the noise suppression coefficient generator 8 included in FIG. The noise suppression coefficient generation unit 81 includes an MMSE STSA gain function value calculation unit 811,
It has a generalized likelihood ratio calculation unit 812, a speech existence probability storage unit 813, and a suppression coefficient calculation unit 814. From the supplied estimated innate SNR and acquired SNR, the point where the suppression coefficient is obtained by calculation is
This is different from the noise suppression coefficient generation unit 8 in FIG. Hereinafter, a calculation method of the suppression coefficient will be described based on the calculation formula described in Reference 1.

The frame number is n and the frequency number is k
And γ_n(k) is supplied from the frequency-dependent SNR calculator 6 in Fig. 1.
Acquired SNR by frequency_n(k) The hat is the estimated destination in Fig. 1.
Estimated innate SN for each frequency supplied from the acquired SNR calculator 7
Let it be R. Also, η_n(k) = ξ_n(k) Hat / q, v_n(k) =
(η_n(k) γ_n(k)) / (1 + η_n(k)). MMSE STSA gay
The function value calculator 811 is the same as the frequency-specific SNR calculator 6
Acquired SNR γ supplied from _n(k), Estimated innate SNR meter in Fig. 1
Estimated innate SNR supplied from calculation unit 7 ξ_n(k) hat and
Voice presence probability supplied from the
MMSE STSA gain function value for each frequency based on the ratio q
Calculated values are output to the suppression coefficient calculation unit 814. For each frequency
MMSE STSA gain function value of G_n(k) is given by equation (18)
Can be In equation (18), I₀(z) is the 0th-order modified Besse
Function, I₁(z) is a first-order modified Bessel function. Deformation
For the Russell function, see "1985, Mathematics Dictionary, Iwanami
Store, page 374.G "(Reference 5).

[0094]

[Equation 17]

The generalized likelihood ratio calculation unit 812 calculates the acquired SNR γ _n (k) supplied from the frequency-specific SNR calculation unit 6 in FIG.
Estimated innate SNR supplied from estimated inferred SNR calculator 7
Based on the voice existence probability q supplied from the R ξ _n (k) hat and the voice existence probability storage unit 813, a generalized likelihood ratio is calculated for each frequency and output to the suppression coefficient calculation unit 814. The generalized likelihood ratio Λ _n (k) for each frequency is given by equation (19).

[0096]

(Equation 18)

[0097] The suppression coefficient calculation unit 814 includes the MMSE STSA gain function value Gn (k) supplied from the MMSE STSA gain function value calculation unit 811 and the generalized likelihood ratio 供給 supplied from the generalized likelihood ratio calculation unit 812. _A suppression coefficient is calculated for each frequency from _n (k) and output to suppression coefficient correction section 15 in FIG. The suppression coefficient G _n (k) bar for each frequency is given by equation (18).

[0098]

[Equation 19]

Instead of calculating the SNR for each frequency, an SNR common to a band composed of a plurality of frequencies can be obtained and used.

Next, as a second configuration example of the frequency-specific SNR calculating section 6, an example of calculating the SNR for each band will be described.

FIG. 12 is a block diagram showing a second example of the configuration of the frequency-specific SNR calculating section 6. As shown in FIG. SNR for each frequency shown in Fig. 32
The difference from the calculation unit 6 is that the band-specific SNR calculation unit 61 has band-specific power calculation units 611 and 612. Band-dependent power calculator 611, the per-band power calculated based on the frequency-noisy speech power spectrum supplied from demultiplexer 602, and outputs it to dividing unit 601 ₀ ~ 601 _K-1. Further, the band-dependent power calculator 612, the per-band power calculated based on the frequency domain estimated noise power spectrum supplied from demultiplexer 603, and outputs it to dividing unit 601 ₀ ~ 601 _K-1.

Next, the configuration and operation of the band-specific power calculator 611 will be described in detail with reference to FIG. FIG. 13 shows a band-by-band power calculation section 611 included in the band-by-band SNR calculation section 61.
FIG. 3 is a block diagram showing the configuration of FIG. Here, the bandwidth L
An example of equal division into M bands with Here, it is assumed that L and M are natural numbers that satisfy the relationship of K = LM.

The band-specific SNR calculation section 61 includes adders 6110 ₀ to
Has 6110 _M-1 . The degraded audio power spectrum for each frequency 910 ₀ to 910 _K-1 (910
_{0 to} 910 _ML-1 ) are adders 6110 ₀ to
Transmitted to 6110 _M-1 . For example, since the frequency number corresponding to the band number 0 is from 0 to L-1, the degraded voice power spectrum 910 ₀ to 910 _{L-1 for each} frequency is added to the adder 6110 ₀
Is transmitted to The frequency number corresponding to the band number 1 because 2L-1 from L, the frequency noisy speech power spectrum 910 _L ~ 9102 _L-1 is transmitted to the adder 61101. The adders 6110 _{0 to} 6110 _M-1 calculate the sums of the supplied degraded voice power spectra for the respective frequencies, and determine the degraded voice power spectra for each band 911 ₀ to 911 _ML-1 (911 ₀ to 91
1 _K−1 ) is output to the division units 601 ₀ to 601 _K−1 in FIG. The calculation result of each adder is output as a band-specific degraded voice power spectrum for each frequency corresponding to each band number. For example, the calculation result of the adder 6110 ₀ is output as the bandwidth noisy speech power spectrum 911 ₀ ~ 911 _L-1. Further, the calculation result of the adder 61101 is output as the band-specific degraded voice power spectrum 911 _L to 911 _2L-1 . Since the configuration and operation are equivalent to the band-based power calculator 611, the description of the band-based power calculator 612 is omitted.

Here, an example in which the frequency band is equally divided into a plurality of bands has been described. However, the critical band described in “1980, Hearing and Speech, IEICE, pp. 115-118” (Reference 6) is described. How to divide, `` 1983, Multirate Digital Signal Processing
l Processing), 1983, Prentice-Hall Inc., USA ”(Reference 7), and other band division methods can be used.

FIG. 14 is a block diagram showing a second embodiment of the present invention. The difference from FIG. 1 which is a block diagram showing the first embodiment of the present invention is that the estimated noise calculation unit 5 is replaced by the estimated noise calculation unit 52, and that the weighted degraded speech calculation unit 14 It does not exist. Less than,
A detailed operation will be described focusing on these differences.

FIG. 15 is a block diagram showing the configuration of the estimated noise calculator 52 included in the second embodiment of the present invention shown in FIG. The difference from FIG. 5 is a first embodiment are that the frequency domain estimated noise calculator 504 ₀ ~504 _K-1 has been replaced with a frequency domain estimated noise calculator 506 ₀ ~ 506 _K-1, That is, the estimation noise calculation unit 52 does not have the weighted degraded speech power spectrum in the input signal. This is because the frequency-dependent estimated noise calculation units 504 _{0 to} 504 _{K -1} require the frequency-dependent weighted degraded voice power spectrum for the input signal,
This is because the estimated noise calculation units 506 ₀ to 506 _K-1 do not need the frequency-dependent weighted degraded speech power spectrum for the input signal. Hereinafter, with reference to FIG. 16 will be described in detail and differences is estimated noise calculator 506 ₀ ~506 _K-1 of the configuration and operation of the FIG.

FIG. 16 is a block diagram showing the configuration of the frequency-dependent estimated noise calculators 506 ₀ to 506 _K-1 included in the estimated noise calculator 52 of FIG. The difference from FIG. 6, which is the first configuration example, is that the frequency-dependent estimated noise calculation section 506 does not have the frequency-weighted degraded speech power spectrum in the input signal, and that the division section 5041 and the non-linear processing section 5042 and a multiplier 5043. Hereinafter, a detailed operation will be described focusing on these differences.

The dividing section 5041 divides the frequency-dependent degraded voice power spectrum supplied from the separating section 502 in FIG. 15 by the estimated noise power spectrum of one frame before supplied from the estimated noise storage section 5942, and divides the result. Non-linear processing unit
Output to 5042. The nonlinear processing unit 5042 having the same configuration and function as the nonlinear processing unit 1485 shown in FIG. 3 calculates a weight coefficient according to the output value of the division unit 5041, and
Output to 043. The multiplier 5043 is connected to the separation unit 502 in FIG.
Calculates the product of the frequency-dependent degraded voice power spectrum supplied from the non-linear processing unit 5042 and the weighting coefficient supplied from the nonlinear processing unit 5042, and outputs the result to the switch 5044.

The output signal of the multiplier 5043 is equivalent to the frequency-dependent weighted degraded speech power spectrum in the frequency-dependent estimated noise calculator 504 in FIG. That is, the frequency-dependent weighted degraded speech power spectrum can be calculated inside the frequency-dependent estimated noise calculation unit 506. Therefore, as the second embodiment of the present invention, it is possible to omit the weighted degraded speech calculator 14.

FIG. 17 is a block diagram showing a third embodiment of the present invention. 1 is the same as FIG. 1 which is a block diagram showing the first embodiment of the present invention, except for the estimated innate SNR calculating unit 71, and therefore, the detailed operation will be described focusing on this difference.

FIG. 18 is a block diagram showing a configuration of the estimated innate SNR calculating section 71 included in FIG. 33 has an acquired SNR storage unit 702, a suppression coefficient storage unit 703, and multiplex multiplication units 705 and 704, but the estimated a priori SNR calculation unit 71 has an estimated noise storage instead. Department
712, emphasis voice power spectrum storage unit 713, S by frequency
It has an NR calculator 715 and a multiplex multiplier 716. Further, the estimated a priori SNR calculation unit 7 has a suppression coefficient in the input signal, but the estimated a priori SNR calculation unit 71 has an enhanced speech amplitude spectrum and an estimated noise power spectrum instead of the suppression coefficient in the input signal. Below, the estimated innate SNR calculators 7 and 71
The detailed operation will be described focusing on these differences existing between.

The multiplex multiplier 716 is the multiplex multiplier 16 of FIG.
From the enhanced speech amplitude spectrum supplied from
The weighted sound power spectrum is obtained by multiplying by the power and output to the emphasized voice power spectrum storage unit 713. The configuration of the multiplex multiplier 716 is the same as that of the multiplex multiplier 17 already described with reference to FIG. 28, and a detailed description thereof will be omitted. The emphasized voice power spectrum storage unit 713 stores the emphasized voice power spectrum supplied from the multiplexing unit 716, and converts the emphasized voice power spectrum supplied one frame before to the SNR calculation unit for each frequency.
Output to 715. The configuration of frequency-specific SNR calculating section 715 is the same as that of frequency-specific SNR calculating section 6 already described with reference to FIG. The estimated noise storage unit 712 is
The estimated noise power spectrum supplied from the estimated noise calculation unit 5 in FIG. 17 is stored, and the estimated speech power spectrum supplied one frame before is output to the frequency-specific SNR calculation unit 715. The frequency-specific SNR calculation unit 715 calculates the SNR of the enhanced speech power spectrum supplied from the enhanced speech power spectrum storage unit 713 and the SNR of the estimated noise power spectrum supplied from the estimated noise storage unit 712 for each frequency, and It is output to the add-on unit 707.

The output signal of the frequency-dependent SNR calculator 715 and FIG.
The output signals of the multiplex multiplier 705 of FIG. 3 are equivalent. Therefore, as the third embodiment of the present invention, it is possible to replace the estimated innate SNR calculation unit 7 with the estimated innate SNR calculation unit 17.

FIG. 19 is a block diagram showing the configuration of the fourth embodiment of the present invention. The difference from FIG. 1 which is a block diagram showing the first embodiment of the present invention is that the estimated noise calculating unit 5 is provided in the estimated noise unit 52, and the estimated innate SNR calculating unit 7 is provided in the estimated innate SNR calculating unit 71. And that the weighted degraded speech calculator 14 does not exist. The configuration and operation of the estimation noise section 52 are the same as those of the present invention.
This is the same as the configuration and operation of the second embodiment. The configuration and operation of the estimated innate SNR calculation unit 71 are the same as those of the third embodiment of the present invention.
This is the same as the configuration and operation of the embodiment. Therefore, FIG.
The fourth embodiment of the present invention shown in FIG. 1 realizes a function equivalent to that of the first embodiment of the present invention shown in FIG.

FIG. 20 is a block diagram showing the configuration of the fifth embodiment of the present invention. The difference from FIG. 1 which is a block diagram showing the first embodiment of the present invention is that the estimated noise calculation unit 5 is replaced by the estimated noise unit 53 and that the speech detection unit 4 is not present. is there. That is, the configuration is such that a voice detection unit is not required for noise estimation. Hereinafter, a detailed operation will be described focusing on these differences.

FIG. 21 shows an estimated noise calculation unit 53 included in FIG.
FIG. 3 is a block diagram showing the configuration of FIG. Differences between the estimated noise calculator 5 shown in FIG. 5, the frequency domain estimated noise calculator 504 ₀
~ 504 _K-1 is the frequency-based estimated noise calculator 508 ₀ ~ 508 _K-1
And that the estimated noise calculation unit 53 does not have a speech detection flag in the input signal. Referring to FIG. 22, frequency-dependent estimation noise calculators 508 ₀ to 508 _K−1
Will be described in detail.

[0117] Figure 22 is a block diagram showing a configuration of a frequency domain estimated noise calculator 508 ₀ ~ 508 _K-1 included in Figure 21. Differences between the frequency domain estimated noise calculator 504 shown in FIG. 6, and the update determination unit 520 is replaced with update decision unit 522, have a voice detection flag input 508 ₀ ~ 508 _K-1 That is not.

FIG. 23 is a block diagram showing a configuration of the update determination unit 522 included in FIG. Update determination unit 5 shown in FIG.
The difference from 20 is that the logical sum calculating unit 5201 is different from the logical sum calculating unit 5221.
That the update determination unit 522 has the logical NOT circuit 5
202 and that the input signal does not have a voice detection flag. That is, the update determination unit 522
Does not use the speech detection flag to update the estimated noise.
This is different from the update determination unit 520 in FIG. Logical OR calculator
The 5221 calculates the logical sum of the output value of the comparison unit 5205 and the output value of the comparison unit 5203, and outputs the calculation result to the switch 5044, the shift register 5045, and the counter 5049 in FIG. That is,
The update determination unit 522 always outputs “1” until the count value reaches a preset value, and after that, outputs “1” when the degraded voice power is smaller than the threshold value.

As described with reference to FIG. 7, comparing section 5205 determines whether or not the degraded voice signal is noise. In other words, it can be said that comparison section 5205 performs voice detection for each frequency. Therefore, it is possible to realize an update determination unit that does not have a voice detection flag as an input.

In all the embodiments described so far,
Although the minimum mean square error short-time spectrum amplitude method has been assumed as a method for removing noise, it can be applied to other methods. An example of such a method is “1979
December, Proceedings of the IEE
Eee, Vol. 67, No. 12 (PROCEEDINGS OF THE IEE
E, VOL.67, NO.12, PP.1586-1604, DEC, 1979), 1586 ~
Page 1604 ”(Literature 8), and the like, but the description of specific examples of these methods will be omitted.

[0121]

As described above, according to the present invention, the power spectrum of noise is estimated using the weighted degraded speech power spectrum. Therefore, it is possible to accurately estimate the power spectrum of noise regardless of the nature of the noise. This makes it possible to obtain an emphasized voice with little distortion and noise.

Further, according to the present invention, noise suppression is performed using the suppression coefficient corrected according to the value of SNR.
It is possible to obtain an emphasized voice with little distortion and noise relative to the SNR value.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a first embodiment of the present invention.

FIG. 2 is a block diagram illustrating a configuration of a weighted degraded speech calculation unit included in the first embodiment of the present invention.

FIG. 3 is a block diagram illustrating a configuration of a multiple nonlinear processing unit included in a weighted degraded speech calculation unit.

FIG. 4 is a diagram illustrating an example of a nonlinear function in a nonlinear processing unit.

FIG. 5 is a block diagram illustrating a configuration of an estimated noise calculator included in the first embodiment of the present invention.

FIG. 6 is a block diagram illustrating a configuration of a frequency-based estimated noise calculator included in FIG. 5;

FIG. 7 is a block diagram illustrating a configuration of an update determination unit included in FIG. 6;

FIG. 8 is a block diagram illustrating a second configuration example of the frequency-based estimated noise calculator included in FIG. 5;

FIG. 9 is a block diagram illustrating a configuration of a suppression coefficient correction unit included in the first embodiment of the present invention.

FIG. 10 is a block diagram illustrating a configuration of a frequency-dependent suppression coefficient correction unit included in FIG. 10;

FIG. 11 is a block diagram illustrating a second configuration example of the noise suppression coefficient generation unit.

FIG. 12 is a block diagram illustrating a second configuration example of the frequency-specific SNR calculation unit.

FIG. 13 is a block diagram illustrating a configuration of a band-specific power calculator included in FIG.

FIG. 14 is a block diagram showing a second embodiment of the present invention.

FIG. 15 is a block diagram illustrating a configuration of an estimated noise calculation unit included in the second embodiment of the present invention.

FIG. 16 is a block diagram illustrating a configuration of a frequency-based estimated noise calculator included in FIG.

FIG. 17 is a block diagram showing a third embodiment of the present invention.

FIG. 18 is a block diagram illustrating a configuration of an estimated innate SNR calculator included in the third embodiment of the present invention.

FIG. 19 is a block diagram showing a fourth embodiment of the present invention.

FIG. 20 is a block diagram showing a fifth embodiment of the present invention.

FIG. 21 is a block diagram illustrating an estimated noise calculator included in the fifth embodiment of the present invention.

FIG. 22 is a block diagram showing a configuration of a frequency-based estimated noise calculator included in FIG. 21.

FIG. 23 is a block diagram illustrating a configuration of an update determination unit included in FIG.

FIG. 24 is a block diagram showing a configuration of a conventional example.

FIG. 25 is a block diagram showing a configuration of a voice detection unit included in the configuration of the conventional example.

FIG. 26 is a block diagram showing a configuration of a power calculator included in FIG. 25.

FIG. 27 is a block diagram illustrating a configuration of a weighted addition unit included in FIG. 25.

FIG. 28 is a block diagram showing a configuration of a multiplex multiplier included in a configuration of a conventional example.

FIG. 29 is a block diagram showing a configuration of an estimated noise calculator included in the configuration of the conventional example.

30 is a block diagram illustrating a configuration of a frequency-based estimated noise calculator included in FIG. 29.

FIG. 31 is a block diagram illustrating a configuration of an update determination unit included in FIG. 30.

FIG. 32 is a block diagram illustrating a configuration of a frequency-specific SNR calculator included in the configuration of the conventional example.

FIG. 33 is a block diagram illustrating a configuration of an estimated innate SNR calculation unit included in the configuration of the conventional example.

FIG. 34 is a block diagram illustrating a configuration of a multiple value range limiting processing unit included in FIG. 33.

FIG. 35 is a block diagram illustrating a configuration of a multi-weighted addition unit included in FIG. 33.

FIG. 36 is a block diagram illustrating a configuration of a noise suppression coefficient generation unit included in the configuration of the conventional example.

FIG. 37 is a block diagram illustrating a configuration of a suppression coefficient search unit included in the noise suppression coefficient generation unit.

[Explanation of symbols]

1 Frame division unit 2 Windowing unit 3 Fourier transform unit 4 Voice detection unit 5, 51, 52, 53 Estimated noise calculation unit 6, 61, 715, 1402 Frequency-specific SNR calculation unit 7, 71 Estimated innate SNR calculation unit 8 , 81 Noise suppression coefficient generation unit 9 Inverse Fourier transform unit 10 Frame synthesis unit 11 Input terminal 12 Output terminal 13, 5049 Counter 14 Weighted degraded speech calculation unit 15 Suppression coefficient correction unit 16, 17, 704, 705, 716, 1404 Multiplex Multiplication unit 401, 1593, 5204, 5206 Threshold storage unit 402, 1594, 5203, 5205 Comparison unit 404, 4075 Constant multiplier 405 Logarithmic calculation unit 406 Power calculation unit 407, 5071, 7071 ₀ to 7071 _K-1 Weighted addition unit 408, 706, 5072 Weight storage unit 409, 5202 Logical NOT circuit 502, 505, 602, 603, 802, 803, 1495, 1502, 1503, 17
02, 1703, 4061, 6111,7013,7072, 7074 Separation section 503, 604, 804, 1475, 1504, 1704, 6115, 7014, 7075
Multiplexer 504 ₀ to 504 _K-1 , 506 ₀ to 506 _K-1 , 507, 508 ₀ to 508
_K-1 , 514 ₀ to 514 _K-1 Estimated noise calculation unit for each frequency 520, 521, 522 Update judgment unit 601 ₀ to 601 _K-1 , 5041, 5048 Division unit 611, 612 Power calculation unit for each frequency 701 Multi-value range limited Processing unit 702 Acquired SNR storage unit 703 Suppression coefficient storage unit 707 Multiple weighted addition unit 708, 4063, 4072, 4074, 5046, 6110 ₀ to 6110 _M-1 adder 712, 1401, 5942 Estimated noise storage unit 713 Enhanced voice Power spectrum storage unit 801 ₀ to 801 _K-1 suppression coefficient search unit 811 MMSE STSA Gain function value calculation unit 812 Generalized likelihood ratio calculation unit 813 Voice existence probability storage unit 814 Suppression coefficient calculation unit 901 Degraded speech power 902 Threshold value 903,923 Weight 904 Update threshold value 905 Weighted adder control signal 910 ₀ to 910 _K-1 , 910 ₀ to 910 _ML-1 Degraded voice power spectrum by frequency 911 ₀ to 911 _K-1 , 911 ₀ to 911 _ML-1 Degraded voice by band power spectrum 921 instantaneous estimated SNR 921 ₀ ~ 921 _K-1 frequency-instantaneous estimated SNR 922 past estimated SNR 922 ₀ ~ 922 _K-1 previous frequency Another estimated SNR 924 estimated apriori SNR 924 ₀ ~ 924 _K-1 frequency-estimated apriori SNR 1405 multiplex nonlinear processor _{1485 0 ~ 1485 K-1,} 5042 nonlinear processor 1501 ₀ ~ 1501 _K-1 frequency-suppression coefficient correction Unit 1591, 7012 ₀ to 7012 _K-1 Maximum value selection unit 1592 Suppression coefficient lower limit value storage unit 1595, 5044 Switch 1596 Correction amount storage unit 1597,17010 to 1701 _K-1 , 4062 ₀ to 4062 _K-1 , 4071, 407
3, 5043 Multiplier 5045 Shift register 5047 Minimum value selector 5201, 5211, 5221 Logical OR calculator 5207 Threshold calculator 5941 Register length memory 7011 Constant memory 8011 Suppression coefficient table 8012, 8013 Address converter

──────────────────────────────────────────────────続 き Continued on the front page (51) Int.Cl. ⁷ Identification symbol FI theme coat ゛ (Reference) G10L 9/16 F term (Reference) 5D015 EE05 EE06 5J064 AA01 BA16 BB03 BB07 BB14 BD01 5K052 AA01 DD02 EE12 FF32

Claims (14)

[Claims] An input signal is converted into a frequency domain signal to obtain a frequency domain signal, the frequency domain signal is separated into an amplitude component and a phase component, and a signal-to-noise ratio is calculated using the amplitude component of the frequency domain signal. Determining a suppression coefficient sequentially based on the signal-to-noise ratio, weighting an amplitude component of the frequency domain signal using the suppression coefficient, and an amplitude component of the weighted frequency domain signal and the frequency domain signal. When removing noise by converting the phase component into a time domain signal, the amplitude component of the frequency domain signal is weighted using a correction suppression coefficient obtained by correcting the suppression coefficient. How to remove noise. 2. A correction signal obtained by determining a signal-to-noise ratio using an amplitude component of the frequency-domain signal when sequentially determining a suppression coefficient based on the signal-to-noise ratio, and correcting the signal-to-noise ratio. 2. The method for removing noise according to claim 1, wherein the suppression coefficient is sequentially determined based on a noise-to-noise ratio. 3. When a signal-to-noise ratio is calculated using an amplitude component of a frequency domain signal, noise is estimated using the amplitude component of the frequency domain signal, and the estimated noise and the amplitude component of the frequency domain signal are calculated. 3. The method for removing noise according to claim 1, wherein the signal-to-noise ratio is determined by using the method. 4. When estimating noise using an amplitude component of a frequency domain signal, the amplitude component of the frequency domain signal is weighted to obtain a weighted amplitude component, and noise is estimated using the weighted amplitude component. 4. The method for removing noise according to claim 3, wherein the noise is estimated. 5. An input signal is converted into a frequency domain signal to obtain a frequency domain signal, the frequency domain signal is separated into an amplitude component and a phase component, and a signal-to-noise ratio is calculated using the amplitude component of the frequency domain signal. Determining the suppression coefficient sequentially based on the signal-to-noise ratio, weighting the amplitude component of the frequency domain signal using the suppression coefficient, and the amplitude component of the weighted frequency domain signal and the frequency domain signal. When the noise is removed by converting the phase component of the frequency domain signal into a time domain signal, the amplitude component of the frequency domain signal is weighted and weighted when the signal-to-noise ratio is obtained using the amplitude component of the frequency domain signal. Determining a noise amplitude component, estimating noise using the weighted amplitude component, and obtaining a signal-to-noise ratio using the estimated noise and an amplitude component of the frequency domain signal. Law. 6. The noise according to claim 4, wherein when weighting an amplitude component of the frequency domain signal to obtain a weighted amplitude component, a weight is sequentially obtained using a signal-to-noise ratio. Method of removal. 7. A method for weighting an amplitude component of a frequency domain signal to obtain a weighted amplitude component, sequentially obtaining weights using a signal-to-noise ratio, and correcting the signal-to-noise ratio by processing with a non-linear function. 7. The noise removal method according to claim 6, wherein a weight is obtained, and the amplitude component of the frequency domain signal is weighted using the correction weight. 8. A conversion unit for converting an input signal into a frequency domain signal to obtain a frequency domain signal, and separating the frequency domain signal into an amplitude component and a phase component; and a signal pair using the amplitude component of the frequency domain signal. A signal-to-noise ratio calculation unit for obtaining a noise ratio, a suppression coefficient generation unit for sequentially determining a suppression coefficient based on the signal-to-noise ratio, a suppression coefficient correction unit for correcting the suppression coefficient to obtain a corrected suppression coefficient, and the correction A multiplication unit that weights the amplitude component of the frequency domain signal using a suppression coefficient; and an inverse transform unit that converts the amplitude component of the weighted frequency domain signal and the phase component of the frequency domain signal into a time domain signal. An apparatus for removing noise, comprising: 9. A suppression coefficient generation unit for sequentially determining a suppression coefficient based on a signal-to-noise ratio includes a correction signal-to-noise ratio calculation unit for correcting the signal-to-noise ratio to obtain a correction signal-to-noise ratio. 9. The apparatus for removing noise according to claim 8, wherein: 10. A signal-to-noise ratio calculator for obtaining a signal-to-noise ratio using an amplitude component of a frequency-domain signal includes an estimation noise calculator for estimating noise using the amplitude component of the frequency-domain signal. 10. The apparatus for removing noise according to claim 8, wherein 11. An estimated noise calculator for estimating noise using an amplitude component of a frequency domain signal includes a weighted amplitude component calculator for weighting the amplitude component of the frequency domain signal to obtain a weighted amplitude component. 11. The apparatus for removing noise according to claim 10, wherein: 12. A conversion unit for converting an input signal into a frequency domain signal to obtain a frequency domain signal, and separating the frequency domain signal into an amplitude component and a phase component; and a signal pair using the amplitude component of the frequency domain signal. A signal-to-noise ratio calculator for determining a noise ratio, a suppression coefficient generator for sequentially determining a suppression coefficient based on the signal-to-noise ratio, and a multiplier for weighting the amplitude component of the frequency domain signal using the suppression coefficient And at least an inverse transform unit that converts the weighted amplitude component of the frequency domain signal and the phase component of the frequency domain signal into a time domain signal, wherein the signal-to-noise ratio calculation unit is A noise removal device comprising: an estimated noise calculation unit that estimates noise using a weighted amplitude component obtained by weighting an amplitude component. 13. A weighted amplitude component calculator for weighting an amplitude component of a frequency domain signal to obtain a weighted amplitude component includes a signal to noise ratio calculator for calculating a signal to noise ratio. An apparatus for removing noise according to claim 11 or 12. 14. A weighted amplitude component calculator for weighting an amplitude component of a frequency domain signal to obtain a weighted amplitude component, comprising: a signal to noise ratio calculator for calculating a signal to noise ratio; 14. A non-linear processing unit that calculates a correction weight by processing with a non-linear function.
A device for removing noise according to claim 1.